Research

A little fun project that I did, was to fine-tune an Large Language Model to solve cryptic crosswords. Cryptic crosswords are crossword puzzles that combine elements of normal/definitional crosswords with wordplay elements. A standard clue contains a wordplay part and a definition part, both of which point to a single answer. This makes the answers almost always unique, unlike plain crosswords where there could be multiple possible answers and one needs crossing letters to solve unambiguously. Good clues also tend to trick the solver by integrating the wordplay elements and definitions into a misleading 'surface' reading of the clues. 

An example of a cryptic crossword clue and answer: 

Bird is cowardly, about to fly away. (5)

The answer is RAVEN, which means “bird”, and is = CRAVEN (“cowardly”) - C (the abbreviation for circa or “about”).

Because of the complicated wordplay elements, they are quite challenging for LLMs (and even humans). I managed to get 18% accuracy with T5-small being trained on Guardian Cryptic Crosswords. For comparison, with few-shot learning with a much larger model like ChatGPT-3.5, one could get an accuracy of 10%!

Collider data analysis is an application that neural networks excel at. Colliders produce high dimensional data in very large numbers, and the goal is to detect 'anomalous' events that signal physics beyond the "Standard Model" (SM) of particle physics. Supervised learning frameworks are especially good at classifying data into signal and background after learning from simulated data for particular Beyond Standard Model (BSM) models. Ideally, one should be able to use unsupervised learning methods to learn the SM distribution and detect outliers that could be signal events. This turns out to be difficult due to the high dimensionality of the space and statistical variations within the SM events themselves. This is an active area of research in the particle physics community.

My first project in graduate school was on early universe cosmology. One of the most pressing issues in fundamental physics today is the matter – anti-matter asymmetry in the universe. Matter and anti-matter are almost perfectly symmetric in the Standard Model of particle physics, the best theory we have so far to explain elementary particles and their behaviour. But in our universe, we see only matter in large amounts, and we have not detected any macroscopic amounts of anti-matter. This is puzzling, and there are many candidate models that try to explain this, mostly through phase transitions in the early universe. We studied a class of such models and through numerical simulations, found that it is more difficult to achieve this asymmetry than previously thought. However, there is still some hope that such a mechanism might work with the right parameters in the theory.

Another imperfection with the Standard Model is that it considers neutrinos to be massless. Neutrinos are very light neutral particles that interact extremely weakly with ordinary matter. They were thought to be completely massless, but a number of experiments conducted in the past few decades have conclusively proven that they, in fact, have a very small mass. This means that the Standard Model should be modified to accommodate neutrino masses, which also necessitates the inclusion of (model-dependent) new physics phenomena that might be observable in experiments soon. Besides the masses being non-zero, it is also very interesting that the neutrinos have such a tiny mass compared to other massive particles that we see, like electrons and protons. I worked on "clockwork" neutrino models that seek to explain the small masses of neutrinos, by adding so-called ‘sterile neutrinos’ to the Standard Model in a particular way that naturally makes the observed neutrinos have very light masses.  We also examined ways in which such models can be tested in experiments like the Large Hadron Collider at CERN, Geneva.